Re: + kmod-avoid-deadlock-by-recursive-kmod-call.patch added to -mm tree

[Tejun Heo <tj@kernel.org>]

Hello,

On Mon, Jan 30, 2012 at 02:03:35PM +0100, Oleg Nesterov wrote:
> Perhaps we can use another system_wq, but afaics WQ_UNBOUND makes sense
> in this case. I mean, there is no reason to bind this work to any CPU.
> See also below.

I've been trying to nudge people away from using special wqs or flags
unless really necessary.  Other than non-reentrancy and strict
ordering, all behaviors are mostly for optimization and using them
incorrectly / spuriously usually doesn't cause any visible failure,
making it very easy to get them wrong and if you have enough of wrong
/ unnecessary usages in tree, the whole thing gets really confusing
and difficult to update in the future.

> > Is it expected consume large
> > amount of CPU cycles?
> 
> Currently __call_usermodehelper() does kernel_thread(), this is almost
> all. But it can block waiting for kernel_execve().

Blocking is completely fine on any workqueue.  The only reason to
require the use of unbound_wq is if work items would burn a lot of CPU
cycles.  In such cases, we want to let the scheduler have full
jurisdiction instead of wq regulating concurrency.

> Not sure this really makes sense, but if we kill khelper_wq perhaps we
> can simplify this code a bit. We can change __call_usermodehelper()
> 
> 		if (wait == UMH_WAIT_PROC)
> 	-		pid = kernel_thread(wait_for_helper, sub_info,
> 	-				    CLONE_FS | CLONE_FILES | SIGCHLD);
> 	+		wait_for_helper(...);
> 		else
> 
> IOW, the worker thread itself can do the UMH_WAIT_PROC work. This makes
> this work really "long running", but then we can kill sub_info->complete
> and use flush_work().

Again, long-running in the sense that the work item spending a lot of
time sleeping should be fine on system_wq or any other wq with default
attributes.  AFAICS, the things to consider here are...

* If work items are expected to consume large amount of CPU cycles (as
  in crypto work items), consider using system_unbound_wq / WQ_UNBOUND.

* If per-domain concurrency limit is necessary (ie. the number of
  concurrent work items doing this particular task should be limited
  rather than consuming global system_wq limit), a dedicated workqueue
  would be better.

Thanks.

-- 
tejun